See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220419649

Multibiometric Systems

Article  in  Communications of the ACM · January 2004

DOI: 10.1145/962081.962102 · Source: DBLP

CITATIONS READS
336 338

2 authors, including:

Arun Ross
Michigan State University
310 PUBLICATIONS   16,815 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Makeup-invariant Face Recognition View project

Image Forensics View project

All content following this page was uploaded by Arun Ross on 23 November 2014.

The user has requested enhancement of the downloaded file.

 The latest research indicates using a combination
of biometric avenues for human identification is more
effective, and far more challenging.

Multibiometric
Systems
BY ANIL K. JAIN AND
ARUN ROSS

T
raditionally, passwords (knowledge-based security)
and ID cards (token-based security) have been used
to restrict access to secure systems. However, secu-
rity can be easily breached in these systems when a
password is divulged to an unauthorized user or a
card is stolen by an impostor. Furthermore, simple
passwords are easy to guess by an impostor and difficult pass-
words may be hard to recall by a legitimate user. The emergence
of biometrics has addressed the problems that plague traditional
verification methods. Biometrics refers to the automatic identifi-
cation (or verification) of an individual (or a claimed identity) by
using certain physiological or behavioral traits associated with
the person (see Figure 1). By using biometrics
Illustration by Sandra Dionisi

34 January 2004/Vol. 47, No. 1 COMMUNICATIONS OF THE ACM

COMMUNICATIONS OF THE ACM January 2004/Vol. 47, No. 1 35
 it is possible to establish an identity based on “who
you are,” rather than by “what you possess” (for
example, an ID card) or “what you remember” (for
example, a password). Current biometric systems
make use of fingerprints, hand geometry, iris, retina,
face, facial thermograms, signature, gait, palm print
and voiceprint to establish a person’s identity [4].
While biometric systems have their limitations
they have an edge over traditional security methods
in that they cannot be easily stolen or shared. Besides
bolstering security, biometric systems also enhance
user convenience by alleviating the need to design
and remember passwords. Moreover, biometrics is
one of the few techniques that can be used for nega-
tive recognition where the system determines
whether the person is who he or she denies to be.
Biometric systems can operate in one of two

WHILE biometric systems have their limitations they have an edge over
traditional security methods in that they cannot be easily stolen or shared.
Besides bolstering security, biometric systems also enhance user convenience
by alleviating the need to design and remember passwords. Moreover, biometrics
is one of the few techniques that can be used for negative recognition where
the system determines whether the person is who he or she denies to be.
modes—the identification mode, in which the iden-
tity of an unknown user is determined, and the veri-
fication mode, in which a claimed identity is either
accepted (a genuine user) or rejected (an impostor).
Biometric systems are being deployed in various
applications including computer logins, ATMs, gro-
cery stores, airport kiosks, and driver’s licenses. The
successful installation of biometric systems in these
apºplications does not imply that biometrics is a
solved problem. In fact, there is significant room for
improvement in biometrics as suggested by the error
rates shown in the table on the next page.
Biometric systems installed in real-world applica-
tions must contend with a variety of problems.
Among them are:
Noise in sensed data. A fingerprint with a scar and a
voice altered by a cold are examples of noisy inputs.
Noisy data could also result from defective or improp-
erly maintained sensors (for example, accumulation of
dirt on a fingerprint sensor) and unfavorable ambient
conditions (for example, poor illumination of a user’s
face in a face recognition system). Noisy biometric
data may be incorrectly matched with templates in the
database resulting in a user being incorrectly rejected.

36 January 2004/Vol. 47, No. 1 COMMUNICATIONS OF THE ACM

 Figure 1.
Examples of some
of the biometric
traits used for
authenticating an
individual. (Gait
image taken
from www.
findbiometrics.
com.)

Intra-class variations. The biometric data acquired failure to enroll (FTE) rate associated with using a sin-
from an individual during authentication may be very gle biometric trait. There is empirical evidence that
different from the data used to generate the template about 4% of the population may have poor quality
during enrollment, thereby affecting the matching fingerprints that cannot be easily imaged by some of
process. This variation the existing sensors.
is typically caused by a Test Test Parameter False Reject False Accept
Spoof attacks. An
user who is incorrectly Rate (FRR) Rate (FAR) impostor may attempt
interacting with the Fingerprint FVC 2002* Users mostly in the age 0.2% 0.2% to spoof the biometric
group 20-39
sensor, or when sensor trait of a legitimately
FRVT 2002** Enrollment and test images
characteristics are modi- Face were collected in indoor
10% 1%
enrolled user in order
fied (for example, by environment and
be on different days
could to circumvent the sys-
changing sensors, that Voice NIST 2000*** Text dependent 10-20% 2-5%
tem. This type of attack
is, the sensor interoper- *Fingerprint Verification Competition; bias.csr.unibo.it/fvc2002 is especially relevant
ability problem) during **Face Recognition Vendor Test; www.frvt.org/FRVT2002
***National Institute of Standards and Technology; www.nist.gov/speech/tests/spk/2000
when behavioral traits
authentication. such as signature and
Distinctiveness. While a biometric trait is expected State-of-the-art error rates voice are used. However,
to vary significantly across individuals, there may be associated with fingerprint, face, physical traits like finger-
and voice biometric systems. The
large similarities in the feature sets used to represent accuracy estimates of biometric prints are also susceptible
these traits. Thus, every biometric trait has some the- systems depend on a number of to spoof attacks.
oretical upper bound in terms of its discrimination test conditions.
capability. Multibiometric Systems
Non-universality. While every user is expected to Some of the limitations imposed by unimodal bio-
possess the biometric trait being acquired, in reality it metric systems (that is, biometric systems that rely
is possible for a subset of the users to not possess a par- on the evidence of a single biometric trait) can be
ticular biometric. A fingerprint biometric system, for overcome by using multiple biometric modalities [1,
example, may be unable to extract features from the 2, 6]. Such systems, known as multibiometric sys-
fingerprints of certain individuals, due to the poor tems, are expected to be more reliable due to the
quality of the ridges (see Figure 2a). Thus, there is a presence of multiple, fairly independent pieces of

COMMUNICATIONS OF THE ACM January 2004/Vol. 47, No. 1 37

evidence. These systems are also able to meet the
stringent performance requirements imposed by var-
ious applications. Multibiometric systems address
the problem of non-universality, since multiple traits
can ensure sufficient population coverage. Further-
more, multibiometric systems provide anti-spoofing
measures by making it difficult for an intruder to
simultaneously spoof the multiple biometric traits of
a legitimate user. By asking the user to present a ran-
dom subset of biometric traits, the system ensures a
live user is indeed present at the point of data acqui-
sition. Thus, a challenge-response type of authenti-
cation can be facilitated using multibiometric
systems.
A variety of factors should be considered when
designing a multibiometric system. These include the
choice and number of biometric traits; the level in the
biometric system at which information provided by
multiple traits should be integrated; the methodology
adopted to integrate the information; and the cost
versus matching performance trade-off.
The choice and number of biometric traits is
largely driven by the nature of the application, the
overhead introduced by multiple traits (computa-
tional demands and cost, for example), and the corre-
lation between the traits considered. In a cell phone
equipped with a camera it might be easier to combine
the face and voice traits of a user, while in an ATM
application it might be easier to combine the finger-
print and face traits of the user. A commercial multi-
biometric system called BioID (www.bioid.com)
integrates the face, voice, and lip movement of an
individual.
The information presented by multiple traits may
be consolidated at various levels.1 At the feature
extraction level, the feature sets of multiple modalities
are integrated and a new feature set is generated; the
new feature set is then used in the matching and deci-
1
Apart from the three levels mentioned here, fusion is also possible at the sensor level
and rank level [3].
38 January 2004/Vol. 47, No. 1 COMMUNICATIONS OF THE ACM
sion-making modules of the biometric system. At the
matching score level, the matching scores output by
multiple matchers are integrated. At the decision
level, the final decisions made by the individual sys-
tems are consolidated by employing techniques such
as majority voting.
Although integration at the feature extraction level
is expected to perform better than fusion at the other
two levels, it is not always feasible for a number of rea-
sons. First, most commercial systems do not provide
Figure 2a. The failure-to-access to information at this
enroll (FTE) problem as level. Second, the feature
observed in fingerprints.
The four impressions of a spaces of different biometric
user’s fingerprint shown traits may not be compatible.
here cannot be enrolled by
most fingerprint
For example, it is difficult to
systems, due to the combine the minutiae feature
poor image quality of the set of a fingerprint image with
ridges. Consequently,
alternate methods must be the eigen-coefficients of a face
adopted in order to image. Third, even if the fea-
include this user in the ture sets were compatible,
system.
concatenation might result in
a feature vector with a very large dimensionality lead-
ing to the “curse of dimensionality” problem. Fusion
at the decision level is considered rigid due to the
availability of limited information. In fact, the only
type of information available at this level is an
“Accept” or a “Reject” label in the verification mode,
or the identity of the user in the identification mode.
Due to the reasons stated here, fusion at the
matching score level is usually preferred, as it is rel-
atively easy to access and combine the scores pre-
sented by the different modalities. Note that fusion
at this level is a practical compromise between
fusion at the other two levels. In the context of ver-
ification, two distinct approaches exist for fusion at
this level. In the first approach the fusion is viewed
as a classification problem where a feature vector is
constructed using the matching scores output by the
individual matchers; this feature vector is then clas-
sified into one of two classes: “Accept” (genuine
user) or “Reject” (impostor) [9]. In the second
approach the fusion is viewed as a combination
problem where the individual matching scores are
combined to generate a single scalar score, which is
then used to make the final decision. Our experi-
ments suggest the combination approach performs
better than the classification approach [7]; however,
the approach of choice can vary depending on the
database used for testing.
The multibiometric system we have
designed uses the face, fingerprint, and hand
geometry attributes of a person for recogni-
tion. Our database consists of 100 different
users with each user providing five samples
per biometric. Information from these three
modalities was integrated at the matching
score level using the combination approach
[7]. Fingerprints were represented using
minutiae features, and the output of the fin-
gerprint matcher was a similarity score in
the (0,100) range; face images were repre-
sented using eigen-coefficients, and the out-
put of the face matcher was a distance score;
hand-geometry images were represented by
14 feature values (corresponding to the
lengths and widths of the fingers, as well as
the width of the palm), and the output of
the matcher was a distance score. Prior to combining
the raw scores, a normalization scheme was employed
to transform the face and hand geometry scores into
similarity scores in the (0,100) range. If s1, s2, and s3
represent the normalized scores pertaining to the fin-
gerprint, face, and hand geometry modalities, respec-
tively, then the final score, Sfus, was computed as,
Sfus = w1s1 + w2s2 + w3s3. Here w1, w2 and w3 are the
weights associated with the three traits, and
w1+w2+w3=1. In the first set of experiments, equal
weights were assigned to all the three modalities. The
improved matching performance as observed using
the Receiver Operating Characteristic (ROC) curve,
which plots the Genuine Accept Rate (GAR) against
the False Accept Rate (FAR) at various matching
thresholds, is shown in Figure 2b.
It is essential that different biometric traits be given
different degrees of importance for different users.
This is especially significant when biometric traits of
some users cannot be reliably acquired. For example,
users with persistently dry fingers may not be able to
provide good quality fingerprints. Such users might
experience higher false rejects when interacting with
the fingerprint system. By reducing the weight of the
fingerprint trait and increasing the weights associated
with the other traits, the false reject error rate of these
users can be reduced. The biometric system learns
user-specific parameters by observing system perfor-
mance over a period of time. This will appeal to that
segment of the population averse to interacting with a
system that constantly requests a user to provide mul-
tiple readings of the same biometric. The emphasis is
on tuning the system variables automatically, yet
appropriately, to attain performance gain. We com-
pute user-specific weights by an exhaustive search
Figure 2b. The receivertechnique in which various
operating characteristic
(ROC) curve showing the
sets of weights are tried out on
performance gain when a training set of genuine and
the simple sum rule is impostor scores, and selecting
used to combine the
the weight set that results in
matching scores of face,
fingerprint, and hand the least error [5]. Let w1,i,
geometry traits of 100
users (five samples per
w2,i and w3,i be the weights
user). associated with the ith user in
the database. The algorithm
operates on the training set as follows:
• For the ith user in the database, vary weights w1,i,
w2,i and w3,i over the range (0, 1), with the con-
straint w1,i+ w2,i+w3,i =1. Compute Sfus = w1,is1+
w2,is2+w3,is3. This computation is performed over
all scores (that is, both genuine and impostor
scores) associated with the ith user.
• Choose that set of weights that minimizes the
total error rate. The total error rate is the sum of
the false accept and false reject rates pertaining to
this user.
This technique was tested on a subset of 10 users
who provided biometric data over a period of two
months (approximately 30 samples per user per bio-
metric). Figure 3 illustrates the case where reducing the
face weight improves verification accuracy. Our exper-
COMMUNICATIONS OF THE ACM January 2004/Vol. 47, No. 1 39
 performance, which can be further improved by
employing user-specific biometric weights. User-
specific weights aid in reducing the false reject
rate, thereby enhancing user convenience.
It must be noted that deploying a multibiomet-
ric system introduces some overhead in terms of
computational demands and costs. Therefore, it is
important the cost versus performance trade-off is
carefully studied before deploying these systems.
Researchers from the National Institute of Science
and Technology (NIST) used commercially avail-
able biometric products recently to acquire and test
multibiometric data pertaining to 1,000 users [8].
This is an indication of the increased attention that
multibiometric systems are receiving from the gov-
ernment (for various national identification pro-
grams currently under implementation such as the
US-VISIT program) as well as from researchers (see
Matt Turk’s article in this section). c

Figure 3. Eliminating false rejects by employing user-specific
weights. (a), (c), and (e) are templates, and (b), (d), and (f) are
the corresponding test samples (from the same user). At a
matching threshold of 50, equal weighting of the three
normalized matching scores (for fingerprint, hand geometry,
and face) results in a false reject of this user, while user-specific
weighting (weights of 0.6, 0.2, and 0.2 for fingerprint, hand and
face, respectively) results in a correct acceptance of the user.
imental results indicate that employing user-specific
weights further improves matching performance [5].
References
1. Bigun, E.S., Bigun, J., Duc, B., and Fischer, S. Expert conciliation for
multimodal person authentication systems using Bayesian statistics.
In Proceedings of the International Conference on Audio and Video-
Based Biometric Person Authentication. (Crans-Montana, Switzerland,
Mar. 1997), 291–300.
2. Brunelli, R., and Falavigna, D. Person identification using multiple
cues. IEEE Trans. on Pattern Analysis and Machine Intelligence 12, 10
(Oct. 1995). IEEE, NY, 955–966.
3. Ho, T.K. Hull, J.J., and Srihari, S.N. Decision combination in mul-
tiple classifier systems. IEEE Trans. on Pattern Analysis and Machine
Intelligence 16, 1 (1994), IEEE, NY, 66–75.
4. Jain, A.K., Bolle, R., and Pankanti, S. (Eds.). Biometrics: Personal
Identification in Networked Society. Kluwer, Dordrecht, The Nether-
lands, 1999.
5. Jain, A.K., and Ross, A. Learning user-specific parameters in a multi-
biometric system. In Proceedings of the IEEE International Conference
on Image Processing (Rochester, NY, Sept. 22–25, 2002), 57–60.
6. Kittler, J., Hatef, M., Duin, R., and Matas, J. On combining classifiers.
IEEE Trans. on Pattern Analysis and Machine Intelligence 20, 3 (Mar.
1998). IEEE, NY, 226–239.
7. Ross, A., and Jain, A.K. Information fusion in biometrics. Pattern Recog-
nition Letters 24, 13 (Sept. 2003), 2115–2125.
8. Snelick, R., Indovina, M., Yen, J., and Mink, A. Multimodal biometrics:
Issues in design and testing. In Proceedings of International Conference on
Multimodal Interfaces (Vancouver, B.C., Nov. 5–7, 2003).
9. Verlinde, P., and Cholet, G. Comparing decision fusion paradigms
using k-NN based classifiers, decision trees and logistic regression in a
multi-modal identity verification application. In Proceedings of the Inter-
national Conference on Audio and Video-Based Biometric Person Authen-
tication. (Washington, D.C., Mar. 1999), 188–193.

Conclusion Anil K. Jain (jain@cse.msu.edu) is a University Distinguished

Multibiometric systems alleviate a few of the prob- Professor in the Department of Computer Science and Engineering,
lems observed in unimodal biometric systems. Michigan State University, East Lansing, MI.
Besides improving matching performance, they also Arun Ross (ross@csee.wvu.edu) is an assistant professor in the Lane
Department of Computer Science and Electrical Engineering, West
address the problems of non-universality and spoof- Virginia University, Morgantown, WV.
ing. Multibiometric systems can integrate informa-
tion at various levels, the most popular one being This work was funded in part by the Center for Identification Technology Research
fusion at the matching score level where the scores (CITeR), West Virginia University, under the NSF-IUCRC program.
output by the individual matchers are integrated.
The simple sum rule results in improved matching © 2004 ACM 0002-0782/04/0100 $5.00

40 January 2004/Vol. 47, No. 1 COMMUNICATIONS OF THE ACM

View publication stats